automatic i/o scheduling algorithm selection for parallel file systems
TRANSCRIPT
Automatic I/O Scheduling Algorithm Selection for Parallel File Systems
Francieli Zanon Boito UFRGS/UFSC / INRIA Grenoble
Ramon Nou BSC
This work has received partial funding from the European Union's Horizon 2020 Programme (2014-2020) and from Brazilian Ministry of Science, Technology and Innovation through Rede Nacional de Pesquisa (RNP) under the HPC4E Project (www.hpc4e.eu), grant agreement n° 689772.
Components of the collaboration● AGIOS : Application-guided I/O scheduler for parallel file systems
○ Selects the best scheduler for an entire workload using decision trees○ Parallel file system oriented ○ Requests to files
● I/O analyzer : Automatic I/O scheduler selection through online workload analysis○ Selects the best scheduler for the next possible part of the workload
using markov chains and pattern matching on traces○ Local devices oriented (kernel level)○ Requests to blocks
AGIOS
● I/O scheduling library● Multiple scheduling algorithms● Can be used by any I/O service at file level (so far: PFS data servers)● Local decisions
AGIOS
● Selects best scheduling algorithms● Decision tree● Performance improvements
○ 75% over aIOLi and 38% over SJF○ Increases performance for 64% more situations○ Decreases performance for 89% less situations
● Information from traces (the whole workload)
I/O analyzer● Captures traces and stats (i.e.,
each 5 seconds)● Compares traces with old traces
using DTW and heuristics.● Obtains a match or no, a
performance metric and the I/O scheduler used.
○ Updates history
● If there is enough information selects "best” I/O scheduler.
I/O analyzer● If we do not have a match we
use performance to favour a scheduler.
● For example, execution time of 4 parallel processes doing intensive I/O with different I/O schedulers and dynamic.
Collaboration
● Study how parallel workloads and AGIOS benefit from intra-workload scheduler changes○ How long should we wait before issuing a schedule change?
■ Overhead○ Which information can we capture to guide it?
■ Relative offsets, sizes, timestamps, performance
Armed Bandit approach
● Add a naive algorithm for I/O scheduler selection: Armed Bandit○ Each scheduler has selection probability based on the immediate past
performance○ Performance improvements to OrangeFS comparable to static
algorithms ○ Large tests where one choice is clearly better○ Large time windows (4s)
Armed Bandit approach
● Local decisions
New pattern matching approach
● Adding the pattern matching and prediction of I/O Analyzer (work in progress)
● Trace capture and analysis○ We create a virtual disk where files are sequentially positioned
● Incorporated a markov chain predictor● Initial results: good pattern matching capability
○ 1 pattern generates ~5 patterns○ next steps: pattern compression
Collaboration
● Slow but continuous progress.● Scarce resources :)
○ Sometimes collaboration is easier if we have our main project aligned. But it is not the case.
○ Clusters
